Autonomous AI Agents with Cyclical Self-Improvement: A Framework for Experience Consolidation and Parameter-Efficient Fine-Tuning
Year: 2025
Type: preprint
Abstract: We present a novel framework for autonomous AI agents that achieve cyclical self-improvement through experience consolidation and parameter-efficient fine-tuning, analogous to human learning during sleep. Our approach integrates temporal segmentation between task execution and knowledge consolidation phases, hybrid memory architecture combining episodic buffers with semantic knowledge graphs, and experience-to-dataset conversion pipelines with salience detection. The system employs Low-Rank Adaptation (LoRA) with Elastic Weight Consolidation (EWC) for parameter-efficient model refinement while maintaining ethical constraints through dynamic enforcement mechanisms. Experimental validation demonstrates 18.5% improvement in few-shot adaptation, 85% reduction in catastrophic forgetting, and 94.8% accuracy in memory retrieval with 98.6% context compression. The framework achieves progressive capability enhancement while maintaining safety constraints with 99.2% compliance rate and computational efficiency with 10,000× fewer trainable parameters than full fine-tuning approaches.
Author: Abhishek Chatterjee
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Funders: University of Washington, DeepMind
OpenAlex ID: W4413358234